3 ms·

> But you will find no copy of the logo in the model data. You wont find a copy of a plaintext in a cyphertext. But you can still extract the plaintext from th

> But you will find no copy of the logo in the model data. You wont find a copy of a plaintext in a cyphertext. But you can still extract the plaintext from the cyphertext.

fnordpiglet 3y ago
That’s an example of a two way lossless transformation. The data is certainly encoded in cypher text and is directly retrievable and has no other purpose than to contain the original data. The model you can’t directly retrieve the original data, and it has more purposes than producing the original. It requires you to specifically manipulate it to produce the copyrighted material, and just because copyrighted material was used to train it doesn’t mean it can even reproduce a facsimile. I think a better counter example is mpeg and other lossy formats. But again, the format does nothing but carry the original even if it’s not a perfect reproduction. You can’t use it in any other way. Its expressed intent is the reproduction of the copyrighted material with no modification or improvement or derivation. These models are not trained with the intent or purpose of only producing the copyrighted materials. It requires your specific action to induce the reproductions if it’s even possible, but it generally serves other purposes in all other uses. This is more like a xerox than not - you can certainly violate copyright with a xerox. But the existence of the xerox itself isn’t to violate copyright. It’s for other purposes. The ambiguity obviously comes in that a xerox machine wasn’t built by scanning all documents on earth first. But I think the very act of mixing all the other images and documents together into the model, which again, is just a statistical aggregate of everything that was trained with mushed together, turns it into at worst a derived work that falls under fair use.
nl 3y ago
If you search "Superman Logo" you find actual copies of the Superman logo which are served from Google's cache. If you ask a VFX artist to create the "Superman Logo" with Photoshop they'll do an excellent job. The first one isn't copyright violation because it is fair use. The second maybe if it is redistributed but we don't ban the use of photoshop by artists because they can choose to reproduce copyright things with it.
- gedy 3y ago
  I agree, and I honestly think that a big part of the issue with AI image generation is people just really have a hard time conceiving of a technology that can make such accurate images from a relatively small model like this. "It must have a copy" - but van Gogh didn't make paintings of hot rods or whatever, and you can't copyright style or technique.
  - vidarh 3y ago
    I see so many lay persons and even sometimes people with a CS background describe diffusion models as some sort of magic content addressable data store you "just" look up a bunch of original images in and somehow copy pieces of. These debates would get a whole lot better if more people had a least a very basic understanding of how the training process works.

> But you will find no copy of the logo in the model data. You wont find a copy of a plaintext in a cyphertext. But you can still extract the plaintext from the cyphertext.

fnordpiglet 3y ago

That’s an example of a two way lossless transformation. The data is certainly encoded in cypher text and is directly retrievable and has no other purpose than to contain the original data. The model you can’t directly retrieve the original data, and it has more purposes than producing the original. It requires you to specifically manipulate it to produce the copyrighted material, and just because copyrighted material was used to train it doesn’t mean it can even reproduce a facsimile. I think a better counter example is mpeg and other lossy formats. But again, the format does nothing but carry the original even if it’s not a perfect reproduction. You can’t use it in any other way. Its expressed intent is the reproduction of the copyrighted material with no modification or improvement or derivation. These models are not trained with the intent or purpose of only producing the copyrighted materials. It requires your specific action to induce the reproductions if it’s even possible, but it generally serves other purposes in all other uses. This is more like a xerox than not - you can certainly violate copyright with a xerox. But the existence of the xerox itself isn’t to violate copyright. It’s for other purposes. The ambiguity obviously comes in that a xerox machine wasn’t built by scanning all documents on earth first. But I think the very act of mixing all the other images and documents together into the model, which again, is just a statistical aggregate of everything that was trained with mushed together, turns it into at worst a derived work that falls under fair use.

nl 3y ago

If you search "Superman Logo" you find actual copies of the Superman logo which are served from Google's cache. If you ask a VFX artist to create the "Superman Logo" with Photoshop they'll do an excellent job. The first one isn't copyright violation because it is fair use. The second maybe if it is redistributed but we don't ban the use of photoshop by artists because they can choose to reproduce copyright things with it.

gedy 3y ago
I agree, and I honestly think that a big part of the issue with AI image generation is people just really have a hard time conceiving of a technology that can make such accurate images from a relatively small model like this. "It must have a copy" - but van Gogh didn't make paintings of hot rods or whatever, and you can't copyright style or technique.
- vidarh 3y ago
  I see so many lay persons and even sometimes people with a CS background describe diffusion models as some sort of magic content addressable data store you "just" look up a bunch of original images in and somehow copy pieces of. These debates would get a whole lot better if more people had a least a very basic understanding of how the training process works.